Anomaly Detection via Reconstructed Discriminant Networks Leveraging Generative Adversarial Frameworks
Xiaoru Liu, Shifeng Li, Cheng Yan, Xi Luo · 2024
In image anomaly detection area, due to the unpredictability and irregularity of anomalies, manual labeling of abnormal samples is difficult and costly, so the use of supervised methods to solve such tasks is limited, and unsupervised methods has recently attracted significant attention. To solve the above problems. In this paper, a discriminator in generative adversarial training networks is introduced on the basis of a surface anomaly detection method(DRAEM). Firstly, a synthetic anomaly strategy presented in DRAEM is utilized to create synthetic anomaly images, and two discriminators are added to adversarial network to enhance the ability of anomaly samples location. We did experiments on publicly available datasets MVTec AD, our method has improved performance.